The Drip
Anthropic keeps coming for it. Claude Design landed about a day after Justin had been venting about how hard it is to get AI to integrate with design tools — Canva, Figma, Power BI, all the same struggle: bring your own template, hope the layout works, accept the weird ornamentation when it doesn't.
The pattern: instead of helping AI integrate with the tools you already use, Anthropic is bringing the tool into their environment. Design today. What tomorrow?
Inside The Bottle
It's been fun to crank on AI Pathfinder the last couple of days — our first real two-person build, lots of moving parts. The speed isn't the issue. The issue is that AI-assisted work has no natural stopping point.
Here's the process Kellan ran that's worth stealing.
Instead of working through fixes one by one, he had Claude pull every open issue, PR, and concern from the codebase into a deck and a one-pager. Not to make decisions — just to surface them. We sat down for a 30-minute meeting, knocked out 30-something decisions in one pass, recorded the meeting, then fed the transcript and the deck back to Claude to lock in what was V1, what was V2, what was getting cut.
The V1 bar Claude proposed: "Don't be embarrassed in a presentation." Simple. Useful. That became the gate.
Why this matters: Claude will happily tell you what to work on next. And next. And next. Kellan has a name for what happens when you keep saying yes — AI creep. Features you didn't ask for, that may create future problems to fix. You haven't gained time. You've spent it ahead.
Justin's version of the same trap: an early Pathfinder session where Claude became the number one hype person. "You could take this to the next level, you could build this." Hell yeah, let's do it. The next morning: we have to prune this branch. No focused work happened — just a dopamine loop with a code editor.
The reset is simple, even if you've heard it a thousand times. Priority is a singular word. The most important thing right now is this — and it doesn't make sense to worry about anything else until it's done.
We've been moving fast and feeling scattered. The question that's helped: not "what's next" but "what's the gate?" Beginning, middle, end. Not beginning, middle, middle, middle.
Lab Notes
| ■ | Justin's note: AI never says "there's nothing to do." It will always find another feature, another refactor. Knowing when to stop is the skill — and it's harder now than it used to be. |
| ■ | Kellan's note: The more you build, the more you have to fix. AI creep doesn't add hours to your week. It subtracts them from a future week. |
Latest Article
Remembering the art of redaction
On knowing when to stop building. The companion piece to this week's Inside The Bottle.
What Stopped Our Scroll
| ■ | Anthropic — Claude Design — Anthropic ships an in-environment design tool. Same pattern as Claude Code: bring the work into Claude rather than integrate Claude into the work. |
| ■ | OpenAI — Symphony — Open-source spec for orchestrating Codex agents through issue trackers like Linear. Released as a reference implementation, not a maintained product. |
|
